Data vendors are asking themselves the same questions: How can we use the treasure trove of data we’ve amassed to drive actionable insights to users? And while it’s great to have the data, how do we provide context and signals around that data?

In the years since its founding in the 1970s, FactSet has acquired a vast breadth of datasets, comprising market data and reference data; fundamentals, pricing and ownership data; and ever-growing volumes of alternative data. So the vendor began looking into which technologies would allow it to use all of this data to give it a competitive advantage.

The answer was artificial intelligence, says Lucy Tancredi, senior vice president of cognitive computing at FactSet.

“Right now, we are looking at the ability to predict a company’s credit ratings, upgrades and downgrades, we are constantly creating new signals and have a backlog that we pick from to develop,” says Tancredi, who joined FactSet in 1995.

FactSet is blending machine learning (ML) and natural language processing (NLP) algorithms to unravel patterns in the data and predict future events. Its ML models can determine which data features have the most impact on a certain outcome, and in what combination it is likely to occur.

“We make predictions where we would build a machine learning model, with massive amounts of time series financial data,” says Tancredi. “We can feed into these models, so that we can explore many options to fine-tune these predictions and get the highest level of accuracy possible.”

FactSet’s predictive signals can detect, for example, secondary equity offerings, the candidates that are likely to issue follow-on offerings, and what companies are likely to issue new corporate bonds.

“We even predict things like activism,” says Gene Fernandez, FactSet’s CTO. “And when companies are going to have an activism event.”

FactSet’s developers are also employing techniques like word embedding, an NLP technique that can simplify neural networks and improve sentiment analysis. Word embedding has been injected into services such as Named Entity Recognition (NER), which identifies companies, people, locations, health conditions, drug names, numbers, monetary values, and dates from unstructured or semi-structured documents. NER also allows users to link any document with other FactSet content sets, such as historical prices or fundamental data.

Creating accurate predictions

In early 2021, FactSet launched its predictive signals that generate predictive insights and analytics for the financial services industry under its Actionable Insights. The suite of solutions is powered by FactSet’s ML models, which are fed its datasets.

Firms are looking to leverage contextualized data to better understand the context surrounding a trade or decision, and FactSet wants to be able to predict what will happen next for a trade, a decision, or an organization.

Before turning its attention to predicting index memberships, FactSet was generating predictions specifically for investment bankers. The signal in question can predict which companies are likely to be vulnerable to shareholder activism. This is done through feeding over 30 components into an ML model, including ownership data, fundamental data, and estimate data. It then looks at a variety of factors to decipher who is most likely to create an activism campaign due to how many shares they own of a company, and pinpointing previous impactful activism campaigns, unpicking what specific type of ownership data was involved there, and finding companies whose blend of those data points indicates that they are most likely to top FactSet’s list of organizations to watch.

Nearly two years ago, FactSet unveiled High Impact Transcripts (Hit), a tool that examines the language used in earnings calls and extracts information from texts to identify companies whose stock could be subject to large positive or negative price moves in the next 20 days. Hit uses data both from decades of FactSet CallStreet earnings call transcripts along with historical stock prices to be able to make predictions on companies’ stocks.

“Hit can be used either as an additional factor into a trading signal or for research analysts as an input to a proprietary trading model, or to see if a particular transcript was worth reading more closely,” says Tancredi.

“We use earnings calls and transcripts to predict pricing sentiments,” adds Fernandez. “We have created machine-learning models that sift through that and detect how sentiment is going to relate to price movement.”

Using artificial intelligence to generate predictive analytics is not without challenges and limitations, however.

Data quality is an important thing to consider when building out AI models for predictions. “The quality of the data fed into models is crucial to creating accurate insights,” says Peter Bebbington, director and CTO at consultancy Brainpool AI. “And low-quality or low-population data that isn’t representative of the predictions that a firm is aiming for will lead to a model that could inaccurately forecast future events.”

MarketAxess, Bloomberg, and Man Group all offer predictive signals or AI-powered prediction solutions. “Everyone is competing with each other, and you have to have good predictions,” Bebbington says. FactSet intends to differentiate its signals from other offerings by drawing on the quality of the datasets that are fed into its ML models.

“Although many companies offer signals, a lot of times they are just simple rules-based alerts—a current state, not a future state, so maybe when a price exceeds a threshold or a company releases earnings,” Tancredi says. “Those are valuable, but anyone can make them. They don’t really provide alpha because everyone has access to this type of data.”